A reduced bias GMM-like estimator with reduced estimator ... - Cemmap

A reduced bias GMM-like estimator with reduced estimator dispersion

Jerry Hausman Konrad Menzel Randall Lewis Whitney Newey

The Institute for Fiscal Studies Department of Economics, UCL cemmap working paper CWP24/07

A Reduced Bias GMM-like Estimator with Reduced Estimator Dispersion Jerry Hausman, Konrad Menzel, Randall Lewis, and Whitney Newey∗ Draft, September 14, 2007 2SLS is by far the most-used estimator for the simultaneous equation problem. However, it is now well-recognized that 2SLS can exhibit substantial finite sample (second-order) bias when the model is over-identified and the first stage partial R2 is low. The initial recommendation to solve this problem was to do LIML, e.g. Bekker (1994) or Staiger and Stock (1997). However, Hahn, Hausman, and Kuersteiner (HHK 2004) demonstrated that the “moment problem” of LIML led to undesirable estimates in this situation. Morimune (1983) analyzed both the bias in 2SLS and the lack of moments in LIML. While it was long known that LIML did not have finite sample moments, it was less known that this lack of moments led to the undesirable property of considerable dispersion in the estimates, e.g. the inter-quartile range was much larger than 2SLS. HHK developed a jackknife 2SLS (J2SLS) estimator that attenuated the 2SLS bias problem and had good dispersion properties. They found in their empirical results that the J2SLS estimator or the Fuller estimator, which modifies LIML to have moments, did well on both the bias and dispersion criteria. Since the Fuller estimator had smaller second order MSE, HHK recommended using the Fuller estimator. However, Bekker and van der Ploeg (2005) and Hausman, Newey and Woutersen (HNW 2005) recognized that both Fuller and LIML are inconsistent with heteroscedasticity as the number of instruments becomes large in the Bekker (1994) sequence. Since econometricians recognize that heteroscedasticity is often present, this finding presents a problem. Hausman, Newey, Woutersen, Chao and Swanson (HNWCS 2007) solve this problem by proposing jackknife LIML (HLIML) and jackknife Fuller (HFull) estimators that are consistent in the presence of heteroscedasticity. HLIML does not have moments so HNWCS (2007) recommend using HFull, which does have moments. However, a problem remains. If serial correlation or clustering exists, neither HLIML nor HFull is consistent. ∗ [email protected],

[email protected], [email protected], [email protected]. Newey thanks the NSF for research support.

1

The continuous updating estimator, CUE, which is the GMM-like generalization of LIML, introduced by Hansen, Heaton, and Yaron (1996) would solve this problem. The CUE estimator also allows treatment of non-linear specifications which the above estimators need not allow for and also allows for general nonspherical disturbances. However, CUE suffers from the moment problem and exhibits wide dispersion.1 GMM does not suffer from the no moments problem, but like 2SLS, GMM has finite sample bias that grows with the number of moments. In this paper we modify CUE to solve the no moments/large dispersion problem. We consider the dual formulation of CUE and we modify the CUE first order conditions by adding a term of order T −1 . To first order the variance of the estimator is the same as GMM or CUE, so no large sample efficiency is lost. The resulting estimator has moments up to the degree of overidentification and demonstrates considerably reduced bias relative to GMM and reduced dispersion relative to CUE. Thus, we expect the new estimator will be useful for empirical research. We next consider a similar approach but use a class of functions which permits us to specify an estimator with all integral moments existing. Lastly, we demonstrate how this approach can be extended tothe entire family of Maximum Empirical Likelihood (MEL) Estimators, so these estimators will have integral moments of all orders. In the next section we specify a general non-linear model of the type where GMM is often used. We then consider the finite-sample bias of GMM. We discuss how CUE partly removes the bias but evidence demonstrates the presence of the no moments/large dispersion problem. We next consider LIML and associated estimators under large instrument (moment) asymptotics and discuss why LIML, Fuller and HLIM and HFull will be inconsistent when time series or spatial correlation is present. In the next section we consider the no moment problem for CUE and demonstrate the source of the no moment condition. We then consider a general modification to the CUE objective function that solves the no moment problem. We then give a specific modification to CUE and specify the ”regularized” CUE estimator, RCUE, and we prove that the estimator has moments. Lastly, we investigate or new estimators empirically. We find that they do not have either the bias arising from a large number of instruments (moments) or the no moment problem of CUE and LIML and perform well when heteroscedasticity and serial correlation are present. Thus, while further research is required to determine the optimal parameters of the modifications 1 While

a formal proof of absence of moments of CUE does not exist, empirical investigation e.g. Guggenberger (2005), demonstrate that CUE likely has not moments. In the spherical disturbance case, CUE is LIML, which is known to have no moments.

2

we present, the current estimators should allow improved empirical results in applied research.

1

Setup

Consider a sample of T observations of random vectors (Wt , Zt ) satisfying the conditional moment restriction E[%(Wt , β)|Zt ] = 0 where β is a K-dimensional parameter. E.g. in the linear instrumental variables case with a correctly specified first stage,

and

%(wt , β) = yt − xt β

(1)

¯ ¸ ¯ ∂ ¯ E %(Wt , β)¯ Zt = zt = E [Xt |Zt = zt ] = zt π ∂β

(2)

·

For purposes of estimation, we will consider an M-dimensional vector of unconditional moment restrictions

gt (β) = g(wt , β) = h(zt )%(wt , β) =: ht %(wt , β)

where ht is a K-dimensional vector of functions of zt . The two-step GMM estimator for this problem minimizes ˆ −1 g ¯ (β)0 Ω ¯ (β) Q(β) := T g

where ¯ (β) := g

T 1X gt (β) T t=1

ˆ is some consistent preliminary estimator of the variance covariance matrix of the moments. On and Ω the other hand, the CUE estimator in Hansen Heaton and Yaron (1996) minimizes the criterion −1 ˆ ˜ ¯ (β)0 Ω(β) ¯ (β) Q(β) := T g g

3

where 1 X 1 X ˆ Ω(β) := wT (s, t)gs (β)gt (β)0 = wT (s, t)hs h0t %(ws , β)%(wt , β) T T s,t≤T

s,t≤T

is a consistent estimator of the variance-covariance matrix of the moment vector for an appropriate choice of wT (s, t). In the absence of autocorrelation, wT (s, t) = 1l{s = t}, for cluster sampling wT (s, t) = 1l{Gs = Gt } is an indicator for observations s and t belonging to the same cluster, and under general autocorrelation, wT (s, t) are more general kernel weights (see Newey and West (1987) and Andrews (1991)). For the following, denote ¯ k (β) G

¯ ¯ 1 X ∂ gt (β 0 )¯¯ T ∂βk β 0 =β

:=

t≤T

¯ k (β) C

¯ ¯ ∂ 1 X wT (s, t) gs (β 0 )¯¯ gt (β)0 T ∂βk 0 β =β

:=

s,t≤T

where without loss of much generality, we suppose throughout that the parameter of interest β is a scalar. Also,

%t

2

:= %(wt , β0 )

νkt

:=

ζklt

:=

ξklmt

:=

¯ · ¸ ¯ ∂ ∂ %(wt , β0 ) − E %(wt , β0 )¯¯ Zt = zt ∂βk ∂βk ¯ · ¸ ¯ ∂2 ∂2 %(wt , β0 ) − E %(wt , β0 )¯¯ Zt = zt ∂βk ∂βl ∂βk ∂βl ¯ · ¸ ¯ ∂3 ∂3 %(wt , β0 ) − E %(wt , β0 )¯¯ Zt = zt ∂βk ∂βl ∂βm ∂βk ∂βl ∂βm

Finite-Sample Bias in GMM

Using the notation introduced above, efficient two-step GMM solves ¯ βˆGM M )0 Ω ˆ −1 g ¯ (βˆGM M ) = 0 G( ˆ is a consistent preliminary estimator of the covariance matrix of g ¯ (β0 ). For the purposes of this where Ω section, we will in the following take β to be a scalar, and treat Ω as known so we can ignore the potential

4

bias from using a consistent estimator of the optimal weighting matrix instead.2 Denoting ¯ ¯ 1 X ∂2 0 ¯ g (β ) t ¯ 0 2 T ∂β β =β t≤T ¯ 3 X ¯ 1 ∂ gt (β 0 )¯¯ T ∂β 3 β 0 =β

¯ 2 (β) := G ¯ 3 (β) := G

t≤T

from a second-order expansion of the first-order conditions about β0 , we get the approximation

0 =

h i ¯ 0 )0 Ω ¯ 0) + G ¯ 2 (β0 )0 Ω ¯ 0 )0 Ω ˆ −1 g ˆ −1 G(β ˆ −1 g ¯ (β0 ) + G(β ¯ (β0 ) (βˆGM M − β0 ) G(β i 1 h¯ 0 ˆ −1 0 ˆ −1 ¯ ¯ ¯ + G (β ) Ω g (β ) + 2 G (β ) Ω G(β ) (βˆGM M − β0 )2 + op (|βˆGM M − β0 |2 ) 0 2 0 3 0 0 2

Following the argument in Lemma 3.3 in Rilstone et al. (1996), we can replace (βˆGM M − β0 )2 by the square of the first-order approximation ¯ 0 )0 Ω ˆ −1 g ¯ (β0 ) G(β βˆGM M − β0 = − + Op (T −1 ) =: ψ + Op (T −1 ) 0 −1 ¯ ¯ ˆ G(β0 ) Ω G(β0 ) and obtain an approximation up to terms of Op (T −1 ). Defining ¯ ¸ ¯ ∂l ¯ −E %(Wt , β)¯ Zt = zt ∂β l ·

m ¯ lt (β) :=

for l = 1, 2, 3, we can write √ 2 For

˜ −1 n T (βˆGM M − β0 ) = D ˜ + op (T −1 )

the general case, see Newey and Smith (2004)

5

where

n ˜ = = ˜ D

=

· ¸ ¯ 1 (β0 )0 HΩ−1 H0 % m 1 1 1¯ ¯ 2 (β0 )0 Ω−1 G(β ¯ 0 ) ψ2 √ ¯ (β0 ) + G + √ ν 0 HΩ−1 H0 % + √ G3 (β0 )0 Ω−1 g T T T 2 0 −1 0 ¯ 1 (β0 ) HΩ H % ¯ 2 (β0 )0 H0 Ω−1 H0 m ¯ 1 (β0 ) m 1 1 m √ + √ ν 0 HΩ−1 H0 % + √ + Op (T −1 ) ¯ 1 (β0 )0 H0 Ω−1 H0 m ¯ 1 (β0 ) T T Tm ¯ 1 (β0 )0 HΩ−1 H0 m ¯ 1 (β0 )0 HΩ−1 H0 ν 1 2 m m ¯ 1 (β0 ) √ + ν 0 HΩ−1 H0 ν + √ T T T T ¯ 2 (β)0 HΩ−1 H0 % 1 m 1 0 −1 0 √ +√ + ζ HΩ H % T T T

Expanding the approximation from above in

√1 T

around its limit and keeping terms of order up to

√1 T

,

we obtain √

T (βˆGM M − β0 ) ≈

¯ 1 (β0 ) ¯ 1 (β0 )0 HΩ−1 H0 % ¯ 2 (β0 )0 H0 Ω−1 H0 m 1 m 1 ν 0 HΩ−1 H0 % 1 m √ +√ +√ 2 D D D T T T ¯ 1 (β0 )0 HΩ−1 H0 %ν 0 HΩ−1 H0 m ¯ 1 (β0 ) + m ¯ 1 (β0 )0 HΩ−1 H0 %%0 HΩ−1 H0 m ¯ 2 (β0 ) 1 2m −√ 2 D T

where D :=

¯ 1 (β0 )0 HΩ−1 H0 m m ¯ 1 (β0 ) T

Taking expectations, E[βˆGM M ] − β0

≈

1 T

"

−1 0 ¯ ¯ 1 (β0 )0 HΩ−1 H m1 (β0 ) m tr (Ω−1 Ω%ν ) %% Ω%ν Ω −2 D D2

#

where we denote Ω%% := E[H0 %%0 H|Z] = Ω and Ω%ν := E[H0 %ν 0 H|Z]. In the absence of autocorrelation, as in Theorem 4.5 in Newey and Smith, E[βˆGM M ] − β0

≈ =

£ ¤ £ ¤ ¯ 1 (β0 )m ¯ 1 (β0 )0 HΩ−1 h0t E νt %t ht Ω−1 h0t E νt %t ht Ω−1 H0 m −2 TD T D2 0 E [νt %t ht P1 ht ] κ ≥ (M − 2) TD TD

where P1 = Ω−1/2 (I − 2PΩ−1 H0 m1 )Ω−1/2 and κ := mint

n

E[%t νt |Zt ] σt2

o . Therefore the bias of one-step

GMM using the efficient weighting matrix increases linearly in the number of moments. The CUE estimator is known to remove this bias from the correlation between the moment functions and

6

their Jacobians (see e.g. Donald and Newey (2000), Newey and Smith (2004)), but simulation studies suggest that like LIML in the case of the linear model with spherical errors, it may suffer from a nomoments problem in weakly identified moments (Guggenberger (2006)). Below, we will give a heuristic argument why CUE should be expected to have no integral moments in many common settings and suggest a modification of CUE to fix this problem.

3

Inconsistency of LIML with many Instruments and Clustering

In the following section, we show the inconsistency of LIML for clustered samples in order to illustrate the need for an estimator which remains higher-order unbiased in the presence of autocorrelation. With many weak linear moment conditions, LIML can be thought of as the CUE estimator using an estimator of the covariance matrix which is only consistent under homoskedasticity, "

# T X 1 1 0 2 ˆ Ω(β) := ZZ (yt − xt β) T t=1 T Therefore the expectation of the CUE objective function is equal to ¡ ¢ ¡ ¢ tr E[Zt Zt0 ]−1 Ω(β) (T − 1)E[gt (β)]0 E[Zt Zt0 ]−1 E[gt (β)] tr E[Zt Zt0 ]−1 Ω(β) T −1 ˜ E[Q(β)] ≈ + = Q0 (β) + E[(Yt − Xt β)2 ] E[(Yt − Xt β)2 ] T E[(Yt − Xt β)2 ] where

µ Ω(β) = Var

1 0 Z (y − Xβ) T

¶

1 6= E[(Yt − Xt β)2 ] Z 0 Z T

so that the expectation of the LIML objective function differs from the limiting objective function Q0 (β) by a term which will in general depend on the parameter β except when (Yt , Xt ) are i.i.d. conditional on Zt . As an illustrative example, suppose we have a sample of clusters g = 1, . . . , G where cluster g contains the observations i = 1, . . . , ng , and we want to estimate the linear two-equation model

yig

=

α + xig β + εig

xig

=

zig π + νig

7

using a k-dimensional vector of instrumental variables, zig satisfying

E[εig |Zig = zig ] = E[νig |Zig = zig ] = 0

Denoting σεε := E[ε2ig |Zig ]

σεν := E[εig νig |Zig ]

and for i 6= j, rεε σεε := E[εig εjg |Zig , Zjg ]

rεν σεε := E[εig νjg |Zig , Zjg ]

and assuming that observations are independent across clusters, i.e. for g 6= h,

E[εig εjh |Zig , Zjh ] = E[εig νjh |Zig , Zjh ] = 0

the first-order conditions for LIML with respect to β evaluated at the true parameter (β0 , π0 ) can be written as

σεε

¯ ∂ ˜ ¯¯ Q(β)¯ ∂β β=β0

= =

x0 PZ ε +

X X g,h≤G

Taking expectations,

ε0 PZ ε 0 xε ε0 ε PG εig (zjh π0 + νjh )P(ig,jh) −

Png

g=1

i=1 εig (zig π0 + Png 2 i=1 εig g=1

PG

i≤ng

j≤nh

νig ) X X g,h≤G

εig εjh P(ig,jh)

i≤ng

j≤nh

¯ ¶ G µ ∂ ˜ ¯¯ rεε σεν X X P(ig,jg) Q(β)¯ = rεν − ∂β σεε β=β0 g=1 i6=j

which will in general be different from zero unless the clustering coefficients for ε and that of its interaction with ν are the same. Under large-G/large ng asymptotics LIML is of course still consistent since the minimal eigenvalue converges to zero in probability, but ignoring autocorrelation may lead to inconsistency of LIML under Bekker (1994) many-moment asymptotics. Therefore estimators which implicitly rely on an independence assumption like LIML, JIVE or HLIM (see Hausman Newey and Woutersen, 2007) will likely fail to remove finite-sample bias in common empirical settings exhibiting time-series or spatial correlation and/or clustering since these methods do not remove

8

the correlations between moments and their Jacobians across observations from the objective function. In contrast, CUE with a correctly specified weighting matrix has been shown to remove that component of the bias (Donald and Newey (2000), Newey and Smith (2004)).

4

The No-Moment Problem for CUE

The source of the no-moment problem is that, as we will now show, under weak identification the true ˜ 0 ) > 0 in the positiveparameter β0 does not satisfy the second-order condition for a minimum, ∇ββ Q(β definite matrix sense with probability 1. To see why this will result in poor performance of CUE, consider again the case of linear moment restrictions,

gt (β) = zt (yt − xt β)

for which

h i−1 ˜ βˆCU E ) ˜ βˆCU E ) + op (1) βˆCU E = − ∇ββ Q( ∇β Q(

ˆ will be near singular with ˜ β) If the second-order condition fails with probability greater than zero, ∇ββ Q( some probability,3 resulting in ”extreme” estimates for β. Denoting derivatives with subscripts, in the scalar parameter case the first-order conditions for CUE can be written as

0=

∂ ˜ Q(β) ∂βk

−1 ¯ −1 ¯ −1 ˆ ˆ ˆ ¯ (β)0 Ω(β) ¯ (β)0 Ω(β) ¯ (β) = Tg Gk (β) − T g Ck (β)Ω(β) g n o −1 ¯ −1 ¯ −1 ˆ ˆ ˆ ¯ (β)0 Ω(β) ¯ (β)¯ = Tg Gk (β) − T tr Ω(β) Ck (β)Ω(β) g g(β)0 n n h io £ ¤o −1 −1 ¯ −1 ˆ ˆ ¯ k (β)¯ ¯ k (β) + tr Ω(β) ˆ ˆ ¯ (β)¯ = tr Ω(β) TG g(β)0 − C Ck (β)Ω(β) Ω(β) − T g g(β)0

for each k = 1, . . . K. In order to verify that the second-order condition for a minimum at β0 doesn’t hold with probability 1, ˜ it is sufficient to show that the expectation of the first derivative of ∇ββ Q(β) is singular, or has deficient 3 By

continuity of the eigenvalues and the intermediate value theorem.

9

rank. We can write ˜ β β (β0 ) = tr Q l k

³

ˆ 0 )−1 [B1 + B2 + 2B3 ] + B4 Ω(β

´ (3)

where

B1

=

B2

=

B3

=

B4

=

¯ ¯ 1 X ∂ ∂ 0 0¯ ¯ ¯ T Gl (β0 )Gk (β0 ) − wT (s, t) gs (β) gt (β) ¯ T ∂βl ∂βk β=β0 s,t≤T # " T ¯ ¯ X ∂2 ¯ ¯ 1 X ∂2 ¯ ¯ (β0 ) − g gt (β0 ) gt (β)¯ gs (β)¯¯ wT (s, t) ∂βl ∂βk T ∂βl ∂βk β=β0 β=β0 t=1 s,t≤T £ ¤ £ ¤ ¯ l (β0 )Ω(β ˆ 0 )−1 T G ¯ k (β0 )¯ ¯ k (β0 ) − C ¯ k (β0 )Ω(β ˆ 0 )−1 T G ¯ l (β0 )¯ ¯ l (β0 ) C g(β0 )0 − C g(β0 )0 − C ih i ∂ hˆ −1 ¯ k (β)Ω(β) ˆ ˆ 0) − T g ¯ (β0 )¯ Ω(β)−1 C Ω(β g(β0 )0 = op (1) ∂βl

ˆ where B4 converges in probability to zero by the Slutsky theorem and the assumption that Ω(β) is a consistent estimator of the GMM variance covariance matrix. In the following we will make the usual GMM assumptions (see e.g. Gallant and White, 1988), in particular the moment functions and their first derivatives are assumed to have bounded 4 + δth moments, and the moment functions evaluated at the true parameter have mean zero. Furthermore we will consider two different forms of autocorrelation.

4.1

Case 1: Finitely Correlated Errors

Assumption 1 There is τ0 τ0 and k, l, m, n = 1, . . . , K,

E[%t %t+τ |Zt , Zt+τ ] = E[νkt νl,t+τ |Zt , Zt+τ ] = E[ζklt ζmn,t+τ |Zt , Zt+τ ] = 0

and E[%t νk,t±τ |Zt , Zt±τ ] = E[%t ζkl,t±τ |Zt , Zt±τ ] = E[νkt ζlm,t±τ |Zt , Zt±τ ] = 0 This assumption covers both the no autocorrelation case and clustered sampling. In this case, we can

10

choose weights wT (s, t) = 1l{|s − t| ≤ τ0 } so that ¯  ¯ ¯ X ¯ ¯ 1 ∂ ∂ ¯ Z (1 − wT (s, t)) = E gs (β) gt (β)0 ¯¯ ¯ T ∂βl ∂βk β=β0 ¯ s,t≤T ¯ ¯ · ¸ · ¸ ¯ ¯ 1 X ∂ ∂ 0 ¯ ¯ = %(ws , β0 )¯ Zs = zs E %(wt , β0 )¯ Zt = zt hs ht E T ∂βl ∂βk |s−t|>τ0 ¯   ¯ X ¯ 1 0  +E hs ht νls νkt ¯¯ Z  T ¯ |s−t|>τ0 ¯ ¯ ¸ · ¸ · ¯ ¯ 1 X ∂ ∂ = hs h0t E %(ws , β0 )¯¯ Zs = zs E %(wt , β0 )¯¯ Zt = zt T ∂βl ∂βk 

E [B1 |Z]

|s−t|>τ0

Similarly,  E [B2 |Z]

µ

X

1 = E T

hs h0t

|s−t|>τ0

¯  ¯ ¸ ¶¯ ¯ ¯ ∂2 ζkls %t + E %(ws , β0 )¯¯ Zs = zs %t ¯¯ Z  = 0 ∂βl ∂k ¯ ·

and ¯  ¯ µ · ¸ ¶¯ ¯ ¯ ∂ 0 ¯ l (β0 )Ω(β ˆ 0 )−1 E  1 ¯ Zs = zs %t ¯ Z  h h ν % + E = C %(w , β ) s ks t s 0 t ¯ ¯ T2 ∂βk ¯ |s−t|>τ0 ¯   ¯ µ · ¸ ¶¯ X ¯ ¯ ∂ 0 ¯ k (β0 )Ω(β ˆ 0 )−1 E  1 ¯ Zs = zs %t ¯ Z  = 0 h h ν % + E %(w , β ) −C s ls t s 0 t ¯ ¯ T2 ∂βl ¯ 

E [B3 |Z]

X

|s−t|>τ0

Hence,  ˆ 0) ˜ β β (β0 )|Z] = tr Ω(β E[Q l k

−1

Weak identification means that

1 T

X |s−t|>τ0

· hs h0t E

 ¯ ¯ ¸ · ¸ ¯ ¯ ∂ ∂ %(ws , β0 )¯¯ Zs = zs E %(wt , β0 )¯¯ Zt = zt  ∂βl ∂βk

¯ ¸ · ¯ ∂ E Ht %(Wt , β0 )¯¯ Zt = zt ∂β

is near singular. As an example, the Staiger and Stock (1997) weak instruments asymptotics for the

11

familiar linear specification (1) correspond to Π E[Xt |Zt = zt ] = zt √ T which is the fastest rate in the first stage for which CUE will be well-posed in the limit. n o ˜ 0 ) := Q ˜ β β (β0 ) Therefore, under weak identification the expectation of ∇ββ Q(β l k

k,l 2. In particular it is possible to apply a mixing law of large numbers. Also for every s, t ∈ {1, 2, . . . }, the kernel weights have to satisfy limT wT (s, t) = 1 and wT (t, t ± mT ) = 0 for some sequence mT −→ ∞, mT T

ˆ −→ 0 in order for Ω(β) to be a consistent estimator of Ω(β). Under these conditions, following an

argument completely analogous to the proof of Lemma 6.6 in Gallant and White (1988), 

 ¯ ¯ · ¸ · ¸ X ¯ ¯ ∂ 1 ∂ ˆ 0 )−1 ˜ β β (β0 )|Z] = tr Ω(β (1 − wT (s, t))hs h0t E E[Q %(ws , β0 )¯¯ Zs = zs E %(wt , β0 )¯¯ Zt = zt  l k T ∂βl ∂βk s,t≤T

This suggests that in the limit we should expect CUE to have no integral moments under weak identification if we allow for the moments and their derivatives to be mixing sequences.

5

A Modification of CUE with Finite Moments

With possibly dependent data, we can derive CUE from the dual maximization problem

max π,β

T X t=1

%(πt )

s.t.

T X

πt gˆt (β) = 0, and

t=1

where for CUE %(v) = −

(v − 1)2 2

12

T X t=1

πt = 1

(4)

˜ := g ˜ (β) containing the moment vector and for a positive definite matrix W 1/2 and the T × k matrix g gt (β) in its t-th column, 1/2

˜ gˆt := Wt g As for the blockwise empirical likelihood estimator described in Kitamura and Stutzer (1997) possible choices for W 1/2 are e.g. W 1/2 :=

1 T −τ (1l{|s

− t| < τ })s,t≤T which corresponds to a matrix with the ³ ´ Bartlett kernel function used for the Newey-West variance estimator, orW 1/2 := ng1(t) 1l{s, t ∈ g(t)} s,t≤T

which yields the robust variance matrix in the presence of one-way clustering. In the dual formulation of the corresponding maximization problem, CUE solves the first-order conditions4

"

where

# ∂ ¯ˆ (β) = 0 ˆ t (β) Ω(β)−1 g π ˆt g ∂β t=1

T X

¯ˆ (βˆCU E )Ω(βˆCU E )−1 g ˆt (βˆCU E ) 1−g π ˆt := PT ¯ (βˆCU E )Ω(βˆCU E )−1 g ˆ ˆs (βˆCU E ) s=1 1 − g

which have an interpretation as “empirical probabilities” in the Generalized Empirical Likelihood framework. We propose to solve a modification of the first-order conditions in which we replace π ˆt with π ˜tα := PT

¯ˆ (βˆCU E )Ω(βˆCU E )−1 g ˆt (βˆCU E ) + αT 1−g ¯ˆ (βˆCU E )Ω(βˆCU E )−1 g ˆs (βˆCU E )] [1 + αT − g

s=1

for some sequence αT > 0 with αT = O

¡1¢ T , e.g. αT =

a T

for some constant a > 0, and hold all terms

except g¯(β) fixed at βˆCU E , i.e. solve " = " ∝

# ∂ ¯ˆ (β) π ˜tα gˆt (βˆCU E ) Ω(βˆCU E )−1 g ∂β t=1

T X

T ³ X

¯ˆ (βˆCU E )Ω(βˆCU E ) 1−g

t=1

−1

# ´ ∂ ¯ˆ (β) ˆt (βˆCU E ) + αT ˆ t (βˆCU E ) Ω(βˆCU E )−1 g g g ∂β 0

Therefore we “bias” the estimates of the “empirical probabilities” toward

1 T

(5)

and skews the first-order

conditions towards those for the two-step optimal GMM estimator. This modification could be interpreted as a regularization in the that it penalizes variation in the ”empirical distribution” (in the GEL sense). 4 see

Newey and Smith (2004)

13

We can rewrite these first-order conditions in vector form as 0

0 = (IK ⊗ ιT )

n h i o ˜ g0 W˜ IK ⊗ (1 + αT ) IT − W1/2 g ˜(˜ g0 W˜ g)−1 g ˜0 W1/2 W1/2 G(˜ g)−1 g ˜0 W1/2 ιT

˜ := g ˜ (β) is a T × M matrix consisting of the where ιT is a T -dimensional column vector of ones, g h i0 ˜ = G ˜ 1 (β)0 , . . . , G ˜ K (β)0 is a KT × M matrix containing moment functions for each observation, and G the stacked first derivatives. For the remaining part of the argument, we assume w.l.o.g. W = IT , the T -dimensional identity matrix. If gt (β) is linear in β, we can solve for βˆα

h

=

n h io i−1 0 ˆ −1 g ˜Ω ˆ −1 G ˜ 0 ιT K (IK ⊗ ιT ) IK ⊗ (1 + αT ) IT − g ˜Ω ˜0 G h n h io i 0 ˆ −1 g ˜Ω ˆ −1 g × (IK ⊗ ιT ) IK ⊗ (1 + αT ) IT − g ˜Ω ˜0 G ˜(0)0 ι

The denominator of the new estimator is equal to £ ¤ ˜ α = (IK ⊗ ιT )0 (1 + αT ) IT − g ˜ 0 (IK ⊗ ιT ) ˜ 0 (IK ⊗ ιT ) =: ∆CU E +αT (IK ⊗ ιT )0 GΩ ˜ −1 G ˜ −1 G ∆ ˜Ω−1 g ˜0 GΩ Since the first summand ¤ 0£ ˜ −1 G ˜ 0 (IK ⊗ ιT ) ∆CU E := (IK ⊗ ιT ) IT − g ˜Ω−1 g ˜0 GΩ is almost surely positive semidefinite,5 the eigenvalues of the denominator £ ¤ ˜ α = (IK ⊗ ιT )0 IT − g ˜ −1 G ˜ 0 (IK ⊗ ιT ) + αT (IK ⊗ ιT )0 GΩ ˜ −1 G ˜ 0 (IK ⊗ ιT ) T∆ ˜Ω−1 g ˜0 GΩ 5 Since the matrix in the middle of this expression has the structure PA where P is a projection matrix, and A is symmetric and positive semi-definite, there exists a decomposition A = XΛX0 where X has full rank and is orthonormal and symmetric, and Λ is a non-negative diagonal matrix. Therefore, we can write

PA = PX0 ΛX = X−1 XPX0 ΛX = X−1 (XPXΛ)X This matrix has the same eigenvalues as XPXΛ, which in turn is positive semi-definite since P is an orthogonal projector, and all elements of Λ are non-negative. Since ∆ is a quadratic form corresponding to a positive semi-definite matrix, ∆ ≥ 0 almost surely.

14

are bounded from below by min eig(T αT V) > 0 for any choice of αT > 0, where 0 ˜ −1 ˜ V := (IK ⊗ ιT ) GΩ G (IK ⊗ ιT )

Following a similar line of reasoning as in Lemma 2 in Fuller (1977), we can apply Lemma B from ¡ ¢ Desgracie and Fuller (1972) to show that for any sequence αT = O T1 , the modification of CUE in (5) has all moments up to the degree of overidentification for T sufficiently large. Also, since this estimator solves a convex combination of the first-order conditions for two-step GMM and CUE, it is asymptotically √ equivalent to both estimators to first order. If T αT → 0, the modified estimator will furthermore be second-order equivalent to CUE.

6

Extensions

From this argument, it can be seen that we can alternatively add any function Jβ (β) which is strictly increasing in β at β0 to the first order condition and solve "

# ∂ ¯ˆ (β) + α Jβ (β) = 0 ˆt (βˆCU E ) Ω(βˆCU E )−1 g π ˆt g ∂β T t=1

T X

(6)

for β, where now α > 0 is held fixed. In the simulations we implement (1)

Jβ (β) =

˜ βˆCU E )0 g ˜ (β) G( , T σ%%

and

(2)

Jβ (β) = β

where the first choice mimics the modification proposed by Fuller (1977). By an argument parallel to that in the preceding section it can be seen that any integral moment of the resulting estimator exists for a sufficiently large sample. The idea of regularizing the “empirical probabilities” for CUE can also be extended to the entire family of Generalized Empirical Likelihood (GEL) Estimators, which solve the problem (4) for a general concave function %(v). As shown in Newey and Smith (2004), the GEL estimator solves "

T X t=1

#0 " T #−1 X 0 ˆ π ˆt Gt (β) kt gˆt (β)ˆ g (β) g¯ˆ(β) = 0 t=1

15

where ˆ 0 gˆt (β)) %0 ( λ π ˆt = PT , 0 ˆ 0 ˆ (β)) s s=1 % (λ g ³ ´−1 0 0 0 ˆ = − 1 PT kt gˆt (β)ˆ ˆ and kt = λ0 gˆ% (λ gˆt )+1 . λ g (β) g¯ˆ(β), t t=1 T t [%0 (λ ˆ0 g ˆs )+1] ˆ0g λ ˆs ³ ´ 1 Op √T , so that we can expand π ˆ around λ0 = 0 and obtain

ˆ = Under regularity conditions, λ

  Ã !−1 µ ¶ µ ¶ T X 1 1 1 ˆ 0 gˆt (β) + Op π ˆt ∝ %0 (0) + %00 (0)λ kt gˆt (β)ˆ gt (β)0 gˆt (β) + Op = − 1 − g¯ˆ(β)0 T T t=1 T where without loss of generality we impose the normalizations %0 (0) = %00 (0) = −1. Also, by L’Hospital’s rule

kt

=

lim

vt ,vs

PT →0

+ lim

%0 (vt ) + 1

vt 0 s=1 vs [% (vs )

+ 1]

%00 (vt )vt −(%0 (vt )+1) vt2

hP

T %0 (vs )+1 s=1 vs

vt ,vs →0

=

1 %000 (0) ˆ 0 ¯ λ (gˆ − gˆt ) + Op + T 2T

µ

1 T2

¶

i

%0 (vt )+1 vt hP i2 0 T % (vs )+1 s=1 vs

ˆ 0 gˆt − λ

hP T s=1

%00 (vs )vs −(%0 (vs )+1) 0ˆ λ gs vs2

Plugging this back into the expansion for π ˆt , it can be seen that π ˆtGEL = π ˆtCU E + Op

i

µ + Op

¡1¢ T , so that in

large samples the projection argument made for CUE will be valid for all GEL estimators. It should also be noted that if the third moments of gˆt (β0 ) are zero and the fourth moments are bounded, all GEL estimators are the same as CUE up to second order. The Monte Carlo results in the next section demonstrate that CUE still has some bias for a large ˆ degree of over-identification most of which arises from the estimation of the optimal weighting matrix Ω. Newey and Smith (2004) showed that Maximum Empirical Likelihood (MEL) does not have this bias, so that applying the proposed modification to the MEL estimator should be expected to have even better finite-sample properties than the modified CUE. We leave this topic for future research.

16

1 T2

¶

7

Monte Carlo Study

We performed a simulation study to investigate the properties of the proposed estimators, ”Regularized” CUE (RCUE) defined by the first-order conditions in equation (5), and RCUE2 (specified in equation (6)), and compare their performance with GMM, LIML, and CUE. In particular, we find that RCUE fixes the no-moments problem present in CUE while significantly reducing the overidentification bias of GMM in both the linear and nonlinear cases.

7.1

Baseline Design: Linear Model without Autocorrelation

For the simulation study we followed the heteroskedastic design of Hausman, Newey, Woutersen, Chao, Swanson (2007) in order to provide empirically relevant parameter choices for the severity of the endogeneity and heteroskedasticity. We then introduced autocorrelation into the model in the form of groupwise clustering and further generalized the design to allow for nonlinear models. The baseline design is as follows:

yi

=

x = εi iid

=

xi β + εi

(7)

z1i π + νi ³ ´ p p ρνi + 1 − ρ2 φθ1i + 1 − φ2 θ2i

(8)

iid

2 ), and νi , θ2i ∼ N (0, 1) in order to allow for heteroskedasticity. θ1i ∼ N (0, z1i

The estimators we chose to compare for the linear model simulations are 2-step GMM, LIML, Fuller(c=1), CUE, and the two RCUE estimators. For RCUE, we found that the adjustment term c=1 (labeled RCU E1 and RCU E21 in the tables) performed well in the simulations, although the choice of the optimal constant will be a topic of future research. We also report the results for c=4 (labeled RCU E4 and RCU E24 ) to illustrate a small range of the bias and variance that the proposed estimators can have. These values for c are the most widely used values for the Fuller estimator since respectively they lead to a mean unbiased or a minimum feasible MSE estimator, respectively. In the nonlinear model, we restricted our attention to GMM, CUE, and the two RCUE estimators with c=1 and c=4. In Table 1, we compare the median bias of the several estimators. All of the estimators have biases which increase with the degree of overidentification, but the bias of GMM is much more pronounced as 17

theory predicts, even approaching that of the OLS bias (unreported) for small values of the concentration parameter, CP, (more weakly identified models) and higher degrees of overidentification, M − K. On the other hand, the bias of CUE and LIML (under homoskedasticity) is much smaller, though also increasing in M − K and decreasing in CP. The RCUE estimators perform similarly to CUE, although they do exhibit a slightly larger median bias. In Tables 2 and 3, we see why it is possible for the RCUE estimators to restore the moments of CUE and lead to lower dispersion of the estimation results. In Table 2, for each set of model parameters, the interquartile range for CUE is greater than for the RCUE estimators. In Table 3, the result is qualitatively similar for the nine decile range (5lower CP and higher M − K.

Thus, these results favor the use of

RCUE over CUE. In Table 4, we see that GMM’s mean bias is increasing in the degree of overidentification, M − K, and is in all cases greater than the bias for the RCUE estimators. In addition, the mean bias for the RCUE estimators is also well defined and stable for all simulation parameters while the inconsistent estimates (using the standard sample mean and variance estimators) of the biases of CUE and LIML demonstrate the instability of their first moment, which arises from the ”moments problem” of non-existence of moments for these estimators. The simulated variance and MSE of the RCUE estimators in Tables 5 and 6 further support the claim that RCUE adjusts CUE in such a way to restore its moments and decrease estimator dispersion. Further, the first two moments of the RCUE estimators are comparable to those of the HFUL estimator in HNWCS (2007) (unreported). Other than the apparent inconsistency of LIML as evidenced by its erratic performance, no other substantive differences in relative performance emerge when testing the estimators under heteroskedasticity.

7.2

Linear Model with Clustering

The next set of simulations was generated by drawing G clusters based on the design in equations (7)-(8) where the disturbances (εi , νi ) are generated as p 1 − γν2 vgi p = ρνgi + 1 − ρ2 θgi

iid

νgi = γν vg +

vg , vgi ∼ N (0, 1)

εgi

θgi ∼ N (0, 1)

18

iid

for a cluster correlation coefficient γν = 0, 0.5, g = 1, . . . , G, and i = 1, . . . , ng , the number of observations in the gth cluster. We also allow for clustering in the instrumental variables,

Zgi = γz ζg +

p

1 − γz2 ζgi

choosing parameter values γz = 0, 0.5. Once we allow for clustering in the first stage with G = 20, 50 clusters given the total sample size of T = 400 in Tables 7 and 10, we first note the large median and mean biases arising from many moments persist in GMM, even after correcting for clustering. Also, LIML, which is typically suggested as a solution to the many moments problem has substantial bias under various conditions. On the other hand, in terms of median bias in Table 7, CUE performs quite well, with much less bias than GMM, although it also exhibits a bias which increases in the degree of overidentification. The RCUE estimators have larger bias, but it still remains much smaller than that of GMM. The same results arise in Tables 8 and 9 in terms of the quantiles. Quantiles of the RCUE estimators are smaller than quantiles of CUE in all cases and provide preliminary evidence for the existence of the first two moments under autocorrelation in Tables 10-12. There we find the mean, variance, and MSE of the estimators generally6 existing for the RCUE estimators, but not for LIML or CUE, whose performance further suggests the moments problem. We also tested HLIM and HFUL from HNWCS (2007) as well as CUE with a heteroskedastic consistent covariance matrix, but all three estimators performed similar to LIML and FULL1 in demonstrating sizable bias.

7.3

Nonlinear Specification

For the Monte Carlo study in Tables 13-18, we changed the specification of the second stage of the baseline model to

yi

= exp {xi β} + εi

xi

=

z1i π + νi

6 c=1 and c=4 are not guaranteed to solve the moments problem, and for a few cases, the evidence suggests that c=1 is not optimal for the RCUE2 estimator. Yet, as previously mentioned the optimal choice of c will be a topic of future research.

19

which has a similar structure as the CCAPM specification in Stock and Wright (2000). The results resemble those in the linear case. In Tables 13 and 16, GMM shows substantial median and mean bias increasing with the degree of overidentification while CUE and the RCUE estimators exhibit much less median bias in all cases. In addition, in Tables 16 and 17, the first two moments of the RCUE estimator exist under all cases, while they do not appear to exist for CUE.7 Thus, the nonlinear simulation results suggest the wide array of the problems to which the RCUE estimators can be applied as they are capable of removing the bias present in GMM under quite general conditions while decreasing the dispersion of the CUE estimator to permit the existence of moments.

8

Conclusion

In this paper we attempt to solve two problems: the bias in GMM which increases with the number of instruments (moment condition) and the ”no moment” problem of CUE that leads to wide dispersion of the estimates. To date, much of the discussion of these problems has been under the assumption of homoscedasticity and absence of serial correlation, which limits their use in empirical research. HNWCS (2005) propose a solution to the homoscedasticity problem for the linear model. In this paper we propose a solution to the more general problem for linear and non-linear specifications when both heteroscedasticity and serial correlation may be present. We propose a class of estimators ”regularized CUE” or RCUE that reduce the moment problem and reduce the dispersion of the CUE estimator.

The Monte Carlo

results accord well with the theory and demonstrate that the new RCUE estimator performs better than existing estimators. As topics for future research, we recommend investigation of the optimal value of the parameter in the RCUE estimators and application of the estimators to the class of Maximum Empirical Likelihood estimators, which is likely to decrease estimator bias even more.

References [1] Andrews, D. (1991): Heteroskedasticity and Autocorrelation Consistent Covariance Matrix Estimation, Econometrica 59(4), 817-58 7 The RCUE2 estimator appears to behave contrary to what the theory would predict as evidenced by its erratic variances in Table 17. The choice of the optimal c will be further investigated to test this estimator.

20

[2] Bekker, P. (1994): Alternative Approximations to the Distributions of Instrumental Variables Estimators, Econometrica 62(3), 657-81 [3] Bekker, P., and J. van der Ploeg (2005): Instrumental Variable Estimation Based on Grouped Data, Statistica Neerlandica 59, 239-67 [4] Degracie, J., and W. Fuller (1972): Estimation of the Slope and Analysis of Covariance when the Concomitant Variable is Measured with Error, Journal of the American Statistical Association 67 No.340, 930-7 [5] Donald, S., and W.Newey (2000): A Jackknife Interpretation of the Continuous Updating Estimator, Economics Letters 67, 239-43 [6] Fuller, W. (1977): Some Properties of a Modification of the Limited Information Estimator, Econometrica 45(4), 939-54 [7] Gallant, A., and H. White (1988): A Unified Theory of Estimation and Inference for Nonlinear Dynamic Models, Oxford [8] Guggenberger, P. (2005): Monte-Carlo Evidence Suggesting a No Moment Problem of the Continuous Updating Estimator, Economics Bulletin 3(13), 1-6 [9] Hahn, J., J. Hausman, and G. Kuersteiner (2004): Estimation with Weak Instruments: Accuracy of higher-order bias and MSE approximations, Econometrics Journal 7, 272-306 [10] Han, Ch., and P.Phillips (2006): GMM with Many Moment Conditions, Econometrica 74(1), 147-92 [11] Hansen, L., J. Heaton, and A. Yaron (1996): Finite-Sample Properties of Some Alternative GMM Estimators, Journal of Business and Economic Statistics 14(3), 262-80 [12] Hausman, J., W. Newey, and T. Woutersen (2005): IV Estimation with Heteroskedasticity and Many Instruments, working paper MIT [13] Hausman, J., W. Newey, T. Woutersen, J. Chao, and N. Swanson (2007): Instrumental Variable Estimation with Heteroskedasticity and Many Instruments, working paper MIT 21

[14] Kitamura, Y., and M. Stutzer (1997): An Information-Theoretic Alternative to Generalized Method of Moments Estimation, Econometrica 65(4), 861-74 [15] Morimune, K. (1983): Approximate Distributions of k-Class Estimators when the Degree of Overidentifiability is Large Compared with the Sample Size, Econometrica 51(3), 821-41 [16] Nagar, A. (1959): The Bias and Moment Matrix of the General k-Class Estimators of the Parameters in Simultaneous Equations, Econometrica 27(4), 575-95 [17] Newey, W., and R. Smith (2004): Higher Order Propertiew of GMM and Generalized Empirical Likelihood Estimators, Econometrica 72(1), 219-55 [18] Newey, W., and K. West (1987): A Simple Positive Semi-Definite Heteroskedasticity and Autocorrelation Consistent Covariance Matrix, Econometrica 55(3), 703-08 [19] Rilstone, P., V. Srivastava, and A. Ullah (1996): The Second-Order Bias and Mean-Squared Error of Nonlinear Estimators, Journal of Econometrics 75, 369-95 [20] Staiger, D., and J. Stock (1997): Instrumental Variables Regression with Weak Instruments, Econometrica 65(3), 557-86 [21] Stock, J., and J. Wright (2000): GMM with Weak Identification, Econometrica 68(5), 1055-96

22

23

CP

0.00 8 0.00 8 0.00 8 0.00 8 0.00 16 0.00 16 0.00 16 0.00 16 0.00 32 0.00 32 0.00 32 0.00 32 0.20 8 0.20 8 0.20 8 0.20 8 0.20 16 0.20 16 0.20 16 0.20 16 0.20 32 0.20 32 0.20 32 0.20 32 ***Simulation

1

R2ε2 |z2 RCU E1

RCU E4

RCU E21

5 0.046 0.064 0.059 10 0.070 0.097 0.075 30 0.130 0.151 0.148 50 0.166 0.182 0.182 5 0.022 0.034 0.023 10 0.031 0.056 0.025 30 0.055 0.084 0.053 50 0.097 0.119 0.101 5 0.011 0.019 0.013 10 0.017 0.033 0.010 30 0.026 0.050 0.015 50 0.036 0.057 0.028 5 0.029 0.045 0.070 10 0.054 0.072 0.075 30 0.123 0.142 0.147 50 0.164 0.179 0.190 5 0.008 0.020 0.026 10 0.019 0.040 0.026 30 0.051 0.076 0.051 50 0.087 0.106 0.089 5 0.004 0.010 0.013 10 0.006 0.022 0.007 30 0.013 0.039 0.006 50 0.027 0.053 0.021 results based on 6250 replications.

M 0.038 0.062 0.135 0.174 0.012 0.015 0.043 0.096 0.007 0.005 0.012 0.025 0.036 0.050 0.135 0.181 0.006 0.011 0.042 0.085 0.002 −0.002 0.001 0.018

RCU E24 0.090 0.156 0.237 0.260 0.053 0.103 0.194 0.227 0.030 0.065 0.143 0.182 0.062 0.134 0.228 0.256 0.027 0.085 0.185 0.224 0.013 0.048 0.135 0.180

GM M 0.026 0.044 0.097 0.131 0.009 0.010 0.030 0.068 0.005 0.004 0.010 0.022 0.015 0.029 0.099 0.143 −0.001 0.004 0.029 0.068 −0.002 −0.005 −0.002 0.013

CU E 0.017 0.024 0.067 0.098 0.006 0.009 0.016 0.028 0.005 0.005 0.005 0.005 −0.222 −0.620 −1.806 −2.579 −0.177 −0.426 −1.629 −2.524 −0.093 −0.211 −1.009 −1.881

LIM L 0.052 0.058 0.090 0.115 0.025 0.026 0.034 0.046 0.014 0.014 0.015 0.014 −0.122 −0.320 −0.837 −1.289 −0.124 −0.320 −1.035 −1.602 −0.077 −0.185 −0.814 −1.444

F U LL1

Table 1: Linear Model under Homoskedasticity and Heteroskedasticity - Median Bias ρ = 0.3, T = 400

24

CP

0.00 8 0.00 8 0.00 8 0.00 8 0.00 16 0.00 16 0.00 16 0.00 16 0.00 32 0.00 32 0.00 32 0.00 32 0.20 8 0.20 8 0.20 8 0.20 8 0.20 16 0.20 16 0.20 16 0.20 16 0.20 32 0.20 32 0.20 32 0.20 32 ***Simulation

1

R2ε2 |z2 RCU E1

RCU E4

RCU E21


M 0.548 0.651 0.915 0.994 0.380 0.453 0.674 0.806 0.257 0.295 0.455 0.584 0.689 0.829 1.165 1.250 0.475 0.556 0.831 0.994 0.327 0.356 0.540 0.670

RCU E24 0.410 0.350 0.237 0.193 0.315 0.288 0.216 0.182 0.232 0.223 0.186 0.164 0.589 0.481 0.312 0.240 0.433 0.375 0.284 0.224 0.310 0.280 0.240 0.202

GM M 0.581 0.717 1.034 1.156 0.387 0.462 0.709 0.878 0.260 0.298 0.459 0.592 0.763 0.936 1.313 1.400 0.493 0.576 0.886 1.054 0.332 0.360 0.546 0.682

CU E 0.575 0.692 0.922 1.060 0.374 0.431 0.575 0.682 0.251 0.273 0.336 0.394 2.015 3.734 9.974 13.901 1.038 1.729 6.297 10.368 0.528 0.746 2.292 3.910

LIM L 0.481 0.582 0.792 0.920 0.346 0.399 0.528 0.625 0.242 0.261 0.325 0.377 1.334 2.040 4.222 5.602 0.852 1.292 3.350 4.853 0.497 0.679 1.625 2.392

F U LL1

Table 2: Linear Model under Homoskedasticity and Heteroskedasticity - Interquartile Range ρ = 0.3, T = 400

25

CP

0.00 8 0.00 8 0.00 8 0.00 8 0.00 16 0.00 16 0.00 16 0.00 16 0.00 32 0.00 32 0.00 32 0.00 32 0.20 8 0.20 8 0.20 8 0.20 8 0.20 16 0.20 16 0.20 16 0.20 16 0.20 32 0.20 32 0.20 32 0.20 32 ***Simulation

1

R2ε2 |z2 RCU E1

RCU E4

RCU E21


M 1.527 1.945 2.913 3.224 1.041 1.282 2.131 2.610 0.669 0.793 1.279 1.718 2.076 2.670 3.780 4.187 1.331 1.798 2.965 3.536 0.841 0.974 1.671 2.253

RCU E24 1.093 0.893 0.587 0.467 0.805 0.725 0.534 0.440 0.586 0.553 0.457 0.398 1.693 1.276 0.776 0.613 1.168 0.997 0.699 0.575 0.791 0.738 0.601 0.515

GM M 1.930 2.708 4.886 5.734 1.094 1.383 2.581 3.580 0.676 0.803 1.315 1.816 3.019 4.636 6.289 6.553 1.443 2.049 3.706 4.465 0.862 1.002 1.736 2.370

CU E 1.945 2.808 5.115 6.310 1.030 1.275 2.203 3.062 0.641 0.701 0.918 1.172 14.835 26.192 59.017 78.679 5.671 12.118 37.617 60.802 1.821 3.096 16.167 33.146

LIM L 1.328 1.664 2.394 2.872 0.907 1.093 1.630 2.062 0.613 0.666 0.861 1.071 3.737 4.608 6.834 8.326 2.971 3.895 6.340 7.898 1.605 2.290 5.150 7.011

F U LL1

Table 3: Linear Model under Homoskedasticity and Heteroskedasticity - Nine Decile Range ρ = 0.3, T = 400

26

CP

0.00 8 0.00 8 0.00 8 0.00 8 0.00 16 0.00 16 0.00 16 0.00 16 0.00 32 0.00 32 0.00 32 0.00 32 0.20 8 0.20 8 0.20 8 0.20 8 0.20 16 0.20 16 0.20 16 0.20 16 0.20 32 0.20 32 0.20 32 0.20 32 ***Simulation

1

R2ε2 |z2 RCU E4

RCU E21

0.024 0.050 0.066 0.066 0.107 0.086 0.132 0.173 0.142 0.155 0.192 0.161 0.002 0.020 0.015 0.017 0.053 0.021 0.045 0.098 0.045 0.077 0.126 0.079 0.001 0.010 0.002 0.005 0.026 −0.001 0.008 0.048 −0.004 0.011 0.059 0.004 0.012 0.033 0.094 0.041 0.082 0.099 0.120 0.160 0.145 0.184 0.204 0.188 −0.010 0.005 0.030 0.001 0.034 0.027 0.036 0.086 0.053 0.083 0.124 0.091 −0.008 −0.001 0.005 −0.007 0.012 −0.003 −0.015 0.029 −0.014 0.002 0.047 −0.001 based on 6250 replications.

RCU E1

5 10 30 50 5 10 30 50 5 10 30 50 5 10 30 50 5 10 30 50 5 10 30 50 results

M 0.022 0.048 0.116 0.138 −0.010 −0.005 0.017 0.052 −0.007 −0.011 −0.018 −0.016 0.039 0.050 0.113 0.176 −0.006 −0.009 0.019 0.066 −0.010 −0.018 −0.035 −0.020

RCU E24 0.078 0.155 0.238 0.260 0.041 0.101 0.195 0.228 0.021 0.059 0.143 0.182 0.047 0.133 0.231 0.258 0.016 0.080 0.187 0.226 0.005 0.043 0.135 0.181

GM M 1.3E+011 -2.0E+011 -4.7E+011 1.1E+011 7.2E+009 4.0E+010 -5.3E+010 -3.3E+011 −0.010 −0.015 -1.3E+011 -9.1E+010 -9.4E+010 -1.2E+011 7.5E+011 -3.0E+011 5.2E+011 7.1E+011 -1.4E+011 -3.0E+010 6.6E+010 4.4E+010 7.1E+009 1.0E+010

CU E −0.008 1.309 −0.522 0.142 −0.028 0.097 −0.310 −0.183 −0.008 −0.008 0.016 0.020 1.047 8.845 −1.698 −1.619 −0.534 −11.442 2.021 −8.330 2.291 −1.058 −2.383 −7.190

LIM L 0.043 0.066 0.102 0.133 0.009 0.014 0.023 0.043 0.003 0.003 0.001 −0.003 −0.060 −0.123 −0.189 −0.231 −0.132 −0.230 −0.470 −0.587 −0.106 −0.200 −0.603 −0.888

F U LL1

Table 4: Linear Model under Homoskedasticity and Heteroskedasticity - Mean Bias ρ = 0.3, T = 400

27

CP

0.00 8 0.00 8 0.00 8 0.00 8 0.00 16 0.00 16 0.00 16 0.00 16 0.00 32 0.00 32 0.00 32 0.00 32 0.20 8 0.20 8 0.20 8 0.20 8 0.20 16 0.20 16 0.20 16 0.20 16 0.20 32 0.20 32 0.20 32 0.20 32 ***Simulation

1

R2ε2 |z2 RCU E1

RCU E4

RCU E21


M 0.210 0.321 0.650 0.803 0.103 0.158 0.399 0.587 0.043 0.061 0.167 0.299 0.386 0.570 1.078 1.311 0.192 0.313 0.697 0.973 0.074 0.103 0.300 0.484

RCU E24 0.120 0.075 0.032 0.020 0.062 0.049 0.026 0.018 0.032 0.029 0.019 0.015 0.305 0.161 0.058 0.034 0.137 0.096 0.047 0.030 0.060 0.052 0.034 0.024

GM M 3.8E+026 3.3E+026 3.6E+026 4.7E+026 3.2E+023 1.5E+025 6.7E+026 1.8E+026 0.045 0.068 4.7E+025 3.5E+025 1.2E+027 1.5E+027 7.3E+026 7.4E+026 1.4E+027 4.5E+026 3.8E+026 2.9E+026 2.7E+025 6.5E+024 1.3E+026 6.9E+025

CU E 5.997 9.2E+003 2.2E+003 267.805 0.309 135.774 107.529 320.255 0.039 0.069 3.596 19.187 1.7E+003 7.5E+005 2.6E+005 2.9E+005 722.773 7.6E+005 1.3E+005 1.8E+005 7.0E+004 1.9E+003 7.5E+003 2.3E+005

LIM L 0.163 0.249 0.492 0.685 0.080 0.117 0.254 0.388 0.035 0.043 0.081 0.132 1.260 2.045 5.461 8.709 0.764 1.325 4.214 7.226 0.309 0.541 2.238 4.594

F U LL1

Table 5: Linear Model under Homoskedasticity and Heteroskedasticity - Variance of Estimates ρ = 0.3, T = 400

28

CP

0.00 8 0.00 8 0.00 8 0.00 8 0.00 16 0.00 16 0.00 16 0.00 16 0.00 32 0.00 32 0.00 32 0.00 32 0.20 8 0.20 8 0.20 8 0.20 8 0.20 16 0.20 16 0.20 16 0.20 16 0.20 32 0.20 32 0.20 32 0.20 32 ***Simulation

1

R2ε2 |z2 RCU E1

RCU E4

RCU E21


M 0.211 0.323 0.664 0.822 0.103 0.158 0.400 0.589 0.043 0.061 0.167 0.300 0.387 0.572 1.091 1.342 0.192 0.313 0.697 0.978 0.074 0.103 0.301 0.484

RCU E24 0.126 0.099 0.088 0.088 0.064 0.059 0.064 0.070 0.033 0.032 0.040 0.048 0.308 0.179 0.111 0.101 0.137 0.102 0.082 0.081 0.060 0.054 0.052 0.057

GM M 3.8E+026 3.3E+026 3.6E+026 4.7E+026 3.2E+023 1.5E+025 6.7E+026 1.8E+026 0.046 0.068 4.7E+025 3.5E+025 1.2E+027 1.5E+027 7.3E+026 7.4E+026 1.4E+027 4.5E+026 3.8E+026 2.9E+026 2.7E+025 6.5E+024 1.3E+026 6.9E+025

CU E 5.997 9.2E+003 2.2E+003 267.825 0.310 135.784 107.625 320.288 0.039 0.069 3.596 19.187 1.7E+003 7.5E+005 2.6E+005 2.9E+005 723.059 7.6E+005 1.3E+005 1.8E+005 7.0E+004 1.9E+003 7.5E+003 2.3E+005

LIM L 0.165 0.253 0.502 0.702 0.080 0.117 0.255 0.390 0.035 0.043 0.081 0.132 1.264 2.061 5.497 8.762 0.782 1.377 4.435 7.571 0.321 0.581 2.601 5.383

F U LL1

Table 6: Linear Model under Homoskedasticity and Heteroskedasticity - Mean Square Error ρ = 0.3, T = 400

29

CP M Z Clustering Clusters RCU E1 RCU E4 8 5 No 20 0.036 0.053 8 10 No 20 0.040 0.064 8 15 No 20 0.047 0.063 32 5 No 20 0.003 0.010 32 10 No 20 0.008 0.018 32 15 No 20 0.016 0.023 8 5 Yes 20 0.102 0.131 8 10 Yes 20 0.087 0.122 8 15 Yes 20 0.057 0.080 32 5 Yes 20 0.025 0.043 32 10 Yes 20 0.032 0.053 32 15 Yes 20 0.030 0.046 8 5 No 50 0.034 0.057 8 10 No 50 0.053 0.086 8 15 No 50 0.050 0.085 32 5 No 50 0.010 0.017 32 10 No 50 0.016 0.029 32 15 No 50 0.014 0.033 8 5 Yes 50 0.071 0.100 8 10 Yes 50 0.081 0.120 8 15 Yes 50 0.084 0.122 32 5 Yes 50 0.017 0.030 32 10 Yes 50 0.022 0.044 32 15 Yes 50 0.025 0.052 ***Simulation results based on 12000 replications.

RCU E21 0.019 0.020 0.023 −0.005 −0.003 0.008 0.086 0.063 0.026 0.008 0.015 0.013 0.024 0.037 0.034 0.003 0.004 −0.001 0.062 0.065 0.074 0.005 0.006 0.007

RCU E24 0.018 0.018 0.022 −0.005 −0.003 0.008 0.081 0.058 0.023 0.007 0.013 0.012 0.020 0.032 0.030 0.002 0.003 −0.002 0.056 0.059 0.065 0.004 0.005 0.006

GM M 0.087 0.151 0.191 0.024 0.059 0.091 0.179 0.230 0.249 0.075 0.132 0.166 0.089 0.156 0.187 0.028 0.064 0.089 0.137 0.197 0.228 0.050 0.099 0.130

CU E 0.004 −0.002 0.014 −0.005 −0.003 0.006 0.007 0.012 0.009 0.000 0.002 0.007 0.005 0.002 0.001 0.002 0.003 −0.002 0.008 0.008 0.016 0.003 0.002 0.001

LIM L 0.011 0.017 0.027 −0.001 −0.003 0.004 0.151 0.206 0.228 0.054 0.102 0.132 0.014 0.021 0.021 0.003 0.002 −0.000 0.088 0.139 0.178 0.021 0.054 0.072

Table 7: Linear Model under Groupwise Clustering - Median Bias ρ = 0.3, T = 400, γν = 0.5 F U LL1 0.045 0.050 0.056 0.009 0.006 0.014 0.162 0.210 0.230 0.061 0.106 0.134 0.051 0.053 0.050 0.012 0.011 0.009 0.108 0.150 0.185 0.030 0.061 0.078

30


RCU E21 0.584 0.665 0.616 0.281 0.367 0.427 0.545 0.545 0.469 0.335 0.403 0.389 0.563 0.648 0.704 0.257 0.312 0.355 0.548 0.607 0.620 0.288 0.349 0.400

RCU E24 0.588 0.672 0.621 0.281 0.367 0.428 0.562 0.559 0.473 0.338 0.406 0.391 0.575 0.659 0.717 0.258 0.313 0.355 0.566 0.628 0.636 0.290 0.351 0.403

GM M 0.421 0.366 0.310 0.242 0.233 0.218 0.349 0.256 0.214 0.236 0.201 0.180 0.418 0.350 0.309 0.233 0.227 0.215 0.371 0.297 0.254 0.232 0.213 0.197

CU E 0.606 0.708 0.632 0.281 0.368 0.429 0.698 0.637 0.487 0.344 0.415 0.397 0.598 0.709 0.773 0.259 0.314 0.356 0.649 0.734 0.730 0.291 0.352 0.408

LIM L 0.569 0.676 0.760 0.254 0.271 0.288 0.417 0.324 0.283 0.251 0.221 0.213 0.578 0.683 0.758 0.249 0.267 0.286 0.496 0.454 0.415 0.251 0.251 0.249

Table 8: Linear Model under Groupwise Clustering - Interquartile Range ρ = 0.3, T = 400, γν = 0.5 F U LL1 0.478 0.572 0.636 0.243 0.260 0.276 0.379 0.309 0.273 0.240 0.215 0.209 0.486 0.578 0.641 0.239 0.256 0.274 0.437 0.416 0.389 0.242 0.243 0.242

31


RCU E21 1.736 2.061 1.959 0.745 0.997 1.171 1.664 1.773 1.572 1.022 1.247 1.229 1.597 1.949 2.157 0.681 0.809 0.968 1.605 1.835 1.989 0.804 1.042 1.204

RCU E24 1.837 2.198 2.034 0.746 0.999 1.176 1.929 1.970 1.673 1.068 1.298 1.267 1.763 2.159 2.365 0.685 0.813 0.976 1.889 2.146 2.285 0.819 1.066 1.233

GM M 1.112 0.924 0.792 0.615 0.585 0.539 1.005 0.701 0.549 0.634 0.534 0.454 1.122 0.901 0.801 0.593 0.559 0.534 1.060 0.772 0.654 0.607 0.539 0.497

CU E 2.129 2.663 2.198 0.747 1.003 1.191 1.7E+013 6.795 2.176 1.189 1.528 1.340 2.098 2.863 3.242 0.686 0.816 0.982 5.156 7.541 5.678 0.831 1.107 1.304

LIM L 1.961 2.728 3.689 0.648 0.706 0.755 1.587 1.032 0.817 0.706 0.633 0.564 1.992 2.723 3.418 0.649 0.697 0.745 1.893 1.632 1.435 0.690 0.671 0.662

Table 9: Linear Model under Groupwise Clustering - Nine Decile Range ρ = 0.3, T = 400, γν = 0.5 F U LL1 1.299 1.648 1.912 0.618 0.672 0.718 1.166 0.926 0.767 0.665 0.607 0.548 1.319 1.665 1.930 0.619 0.664 0.707 1.275 1.246 1.189 0.651 0.639 0.635

32

CP M Z Clustering Clusters RCU E1 RCU E4 8 5 No 20 0.008 0.037 8 10 No 20 0.014 0.055 8 15 No 20 0.014 0.034 32 5 No 20 −0.008 0.001 32 10 No 20 −0.007 0.008 32 15 No 20 −0.006 0.007 8 5 Yes 20 0.081 0.120 8 10 Yes 20 0.069 0.115 8 15 Yes 20 0.018 0.057 32 5 Yes 20 −0.003 0.021 32 10 Yes 20 0.004 0.038 32 15 Yes 20 −0.001 0.024 8 5 No 50 0.008 0.041 8 10 No 50 0.032 0.083 8 15 No 50 0.028 0.089 32 5 No 50 −0.001 0.008 32 10 No 50 0.006 0.025 32 15 No 50 −0.000 0.026 8 5 Yes 50 0.056 0.094 8 10 Yes 50 0.066 0.122 8 15 Yes 50 0.066 0.128 32 5 Yes 50 −0.002 0.016 32 10 Yes 50 0.004 0.038 32 15 Yes 50 0.006 0.046 ***Simulation results based on 12000 replications.

RCU E21 −0.027 −0.036 −0.366 −0.019 −0.026 −0.028 0.021 −0.017 −0.046 −0.043 −0.048 −0.047 −0.011 −0.005 −0.019 −0.011 −0.012 −0.024 0.031 0.014 0.011 −0.022 −0.031 −0.032

RCU E24 −0.046 −0.058 −0.073 −0.020 −0.032 −0.031 −0.010 −0.048 −0.081 −0.054 −0.064 −0.058 −0.034 −0.030 −0.047 −0.013 −0.014 −0.027 −0.001 −0.021 −0.024 −0.026 −0.040 −0.041

GM M 0.074 0.147 0.188 0.016 0.054 0.085 0.170 0.229 0.249 0.060 0.128 0.163 0.071 0.149 0.184 0.019 0.059 0.084 0.128 0.194 0.226 0.036 0.092 0.126

CU E -2.2E+012 -2.6E+012 -9.8E+011 −0.021 -2.0E+010 -6.5E+010 -1.3E+013 -6.2E+012 -1.7E+012 -1.2E+012 -8.7E+011 -3.8E+011 -2.3E+012 -3.4E+012 -3.6E+012 1.6E+010 1.8E+010 -9.1E+010 -1.0E+013 -7.3E+012 -6.0E+012 -1.0E+011 -2.4E+011 -5.6E+011

LIM L −0.028 0.136 −0.036 −0.016 −0.015 0.052 0.215 0.227 0.141 0.012 0.082 0.123 −0.244 −0.059 −0.103 −0.011 −0.012 −0.015 −0.292 0.139 0.085 −0.003 0.037 0.049

Table 10: Linear Model under Groupwise Clustering - Mean Bias ρ = 0.3, T = 400, γν = 0.5 F U LL1 0.035 0.047 0.061 −0.003 −0.002 −0.002 0.152 0.206 0.230 0.039 0.095 0.128 0.034 0.049 0.058 0.001 0.002 −0.004 0.102 0.142 0.174 0.014 0.046 0.069

33


RCU E21 0.288 0.408 1.2E+003 0.056 0.107 0.154 0.257 0.297 2.105 0.112 0.162 0.169 0.235 0.342 0.411 0.045 0.067 0.100 0.227 0.299 0.345 0.065 0.111 0.147

RCU E24 0.367 0.527 0.493 0.057 0.191 0.164 0.365 0.411 0.363 0.140 0.212 0.209 0.313 0.461 0.551 0.047 0.071 0.111 0.325 0.423 0.485 0.074 0.135 0.179

GM M 0.132 0.084 0.060 0.036 0.032 0.027 0.116 0.046 0.028 0.041 0.026 0.020 0.125 0.078 0.060 0.033 0.029 0.026 0.117 0.057 0.040 0.036 0.027 0.023

CU E 2.6E+027 3.2E+027 6.1E+026 0.064 1.2E+025 1.7E+025 1.0E+028 4.3E+027 6.6E+026 8.6E+026 2.8E+026 1.2E+026 5.4E+027 7.2E+027 3.7E+027 3.2E+024 1.7E+024 2.9E+025 1.3E+028 5.3E+027 3.3E+027 1.9E+025 9.0E+025 2.1E+026

LIM L 182.247 1.8E+003 66.550 0.065 0.052 61.775 36.471 7.891 39.801 2.451 0.382 0.038 251.221 64.200 311.093 0.041 0.071 0.153 834.605 54.942 38.744 0.139 0.065 2.581

Table 11: Linear Model under Groupwise Clustering - Variance of Estimates ρ = 0.3, T = 400, γν = 0.5 F U LL1 0.162 0.248 0.317 0.036 0.044 0.051 0.130 0.094 0.064 0.048 0.038 0.031 0.164 0.248 0.324 0.036 0.043 0.051 0.152 0.152 0.143 0.041 0.042 0.044

34


RCU E21 0.289 0.409 1.2E+003 0.056 0.108 0.155 0.258 0.298 2.107 0.113 0.165 0.171 0.235 0.342 0.411 0.046 0.068 0.101 0.228 0.299 0.345 0.066 0.112 0.148

RCU E24 0.369 0.530 0.498 0.058 0.192 0.165 0.365 0.413 0.369 0.143 0.216 0.213 0.314 0.462 0.553 0.047 0.071 0.112 0.325 0.423 0.486 0.075 0.137 0.181

GM M 0.137 0.105 0.095 0.036 0.035 0.035 0.145 0.098 0.090 0.045 0.043 0.046 0.130 0.100 0.094 0.034 0.033 0.033 0.134 0.095 0.091 0.037 0.036 0.039

CU E 2.6E+027 3.2E+027 6.1E+026 0.064 1.2E+025 1.7E+025 1.0E+028 4.3E+027 6.6E+026 8.6E+026 2.8E+026 1.2E+026 5.4E+027 7.2E+027 3.7E+027 3.2E+024 1.7E+024 2.9E+025 1.3E+028 5.3E+027 3.3E+027 1.9E+025 9.0E+025 2.2E+026

LIM L 182.248 1.8E+003 66.551 0.066 0.052 61.778 36.517 7.943 39.821 2.452 0.389 0.053 251.280 64.203 311.103 0.041 0.071 0.153 834.690 54.961 38.752 0.139 0.066 2.583

Table 12: Linear Model under Groupwise Clustering - Mean Square Error ρ = 0.3, T = 400, γν = 0.5 F U LL1 0.163 0.250 0.321 0.036 0.044 0.051 0.153 0.136 0.117 0.050 0.047 0.048 0.165 0.250 0.327 0.036 0.043 0.051 0.162 0.173 0.173 0.042 0.044 0.048

Table 13: Nonlinear Model under Homoskedasticity - Median Bias T = 400 ρˆ CP M GM M CU E RCU E1 RCU E4 0.4 12.50 5 0.050 0.006 0.023 0.036 0.4 12.50 10 0.083 0.008 0.030 0.053 0.4 12.50 30 0.115 0.015 0.047 0.061 0.4 12.50 50 0.124 0.019 0.045 0.058 0.4 25.00 5 0.029 0.005 0.014 0.021 0.4 25.00 10 0.056 0.005 0.017 0.032 0.4 25.00 30 0.096 0.011 0.026 0.043 0.4 25.00 50 0.108 0.017 0.034 0.046 ***Simulation results based on 7000 replications.

RCU E21 0.023 0.021 0.035 0.038 0.013 0.012 0.017 0.027

RCU E24 0.012 0.017 0.040 0.041 0.007 0.008 0.017 0.027

Table 14: Nonlinear Model under Homoskedasticity - Interquartile Range T = 400 ρˆ CP M GM M CU E RCU E1 RCU E4 0.4 12.50 5 0.165 0.239 0.205 0.184 0.4 12.50 10 0.127 0.272 0.216 0.170 0.4 12.50 30 0.080 0.304 0.236 0.169 0.4 12.50 50 0.067 0.298 0.228 0.165 0.4 25.00 5 0.126 0.161 0.149 0.139 0.4 25.00 10 0.105 0.183 0.157 0.134 0.4 25.00 30 0.073 0.226 0.186 0.141 0.4 25.00 50 0.062 0.237 0.195 0.147 ***Simulation results based on 7000 replications.

RCU E21 0.188 0.223 0.259 0.255 0.145 0.166 0.208 0.216

RCU E24 0.222 0.252 0.272 0.262 0.157 0.178 0.217 0.221

Table 15: Nonlinear Model under Homoskedasticity - Nine Decile Range T = 400 ρˆ CP M GM M CU E RCU E1 RCU E4 0.4 12.50 5 0.467 0.800 0.638 0.572 0.4 12.50 10 0.341 0.954 0.695 0.552 0.4 12.50 30 0.205 1.369 0.860 0.572 0.4 12.50 50 0.167 1.388 0.883 0.606 0.4 25.00 5 0.342 0.467 0.421 0.388 0.4 25.00 10 0.280 0.543 0.456 0.385 0.4 25.00 30 0.186 0.786 0.611 0.443 0.4 25.00 50 0.154 0.941 0.730 0.513 ***Simulation results based on 7000 replications.

35

RCU E21 0.483 0.583 0.806 0.915 0.385 0.454 0.649 0.766

RCU E24 0.656 0.774 0.991 1.040 0.440 0.518 0.731 0.837

Table 16: Nonlinear Model under Homoskedasticity - Mean Bias T = 400 ρˆ CP M GM M CU E RCU E1 RCU E4 0.4 12.50 5 0.043 −0.024 0.015 0.029 0.4 12.50 10 0.083 −0.028 0.034 0.057 0.4 12.50 30 0.118 −0.074 0.054 0.073 0.4 12.50 50 0.127 −0.094 0.048 0.066 0.4 25.00 5 0.022 −0.007 0.003 0.012 0.4 25.00 10 0.055 −0.009 0.011 0.029 0.4 25.00 30 0.098 −0.042 0.017 0.043 0.4 25.00 50 0.111 −0.057 0.028 0.047 ***Simulation results based on 7000 replications.

RCU E21 0.023 0.044 0.015 0.038 0.006 0.005 0.000 0.015

RCU E24 0.001 0.034 0.025 0.042 −0.006 −0.004 −0.003 0.013

Table 17: Nonlinear Model under Homoskedasticity - Variance of Estimates T = 400 ρˆ CP M GM M CU E RCU E1 RCU E4 0.4 12.50 5 0.023 0.245 0.048 0.042 0.4 12.50 10 0.011 0.328 0.051 0.036 0.4 12.50 30 0.004 0.271 0.068 0.036 0.4 12.50 50 0.003 0.675 0.070 0.035 0.4 25.00 5 0.012 0.144 0.020 0.017 0.4 25.00 10 0.008 0.167 0.024 0.017 0.4 25.00 30 0.003 0.107 0.041 0.021 0.4 25.00 50 0.002 0.495 0.055 0.026 ***Simulation results based on 7000 replications.

RCU E21 0.022 2.245 1.013 0.408 0.015 0.020 0.042 0.364

RCU E24 0.040 2.901 0.986 0.229 0.020 0.028 0.052 0.154

Table 18: Nonlinear Model under Homoskedasticity - Mean Square Error T = 400 ρˆ CP M GM M CU E RCU E1 RCU E4 0.4 12.50 5 0.025 0.245 0.048 0.043 0.4 12.50 10 0.018 0.329 0.052 0.039 0.4 12.50 30 0.018 0.276 0.071 0.041 0.4 12.50 50 0.019 0.684 0.073 0.039 0.4 25.00 5 0.012 0.144 0.020 0.017 0.4 25.00 10 0.011 0.167 0.025 0.018 0.4 25.00 30 0.013 0.109 0.041 0.023 0.4 25.00 50 0.015 0.498 0.056 0.028 ***Simulation results based on 7000 replications.

36

RCU E21 0.023 2.246 1.013 0.410 0.015 0.020 0.042 0.364

RCU E24 0.040 2.903 0.986 0.231 0.020 0.028 0.052 0.154